●builderYou can implement segmentation capabilities using existing MLLMs without the overhead of training dedicated mask decoders.
●researcherThis demonstrates that text-prompted segmentation can be achieved through geometric refinement rather than specialized architectures.